Roberto Moreno-Díaz Instituto Universitario de Ciencias y Tecnologías Cibernéticas Universidad de Las Palmas de Gran Canaria [email protected] Neurocybernetics and Artificial Intelligence 1. Objective This lecture explores the evolution and content of Neurocybernetics, its relation to Artificial Intelligence and their interrelations with Computer Science and Knowl- edge Engineering. We want to emphasize throughout the presentation the need to return in a system- atic way to the original writings of the main authors, where one can find many of the basic ideas, which are still pending to be developed and which are a source of inspiration. We also want to emphasize the educational potentials these matters of- fer for computing research curricula. In dealing with the relations of Neurocybernetics to Artificial Intelligence, we shall make generous use of some ideas shared with our former and very good friend and colleague, Prof. Jos´e Mira Mira, who passed away unexpectedly and too early. What is presented here is dedicated to his Memory, and to his lively, strong and serious way to make cybernetic and computing science. 156 Roberto Moreno-Díaz 1. The Original Components of Cybernetics Neurocybernetics took off in the Forties although many of the basic ideas had been being managed in philosophic and scientific circles since the times of the Ancient Greeks. From 1943 to 1945, a kind of synergetic process was started up, triggered as the result of three basic pieces of work: Norbert Wiener, Arthur Rosemblueth and Julian Bigelow’s study (1943) on the nature of teleological processes where the crucial idea was that the relevant in a homeostatic process is the information return and not the energy return via the feedback loops (see figure 1). Following this, came the work of the young British philosopher, Kenneth Craick, published in the form of a small book called On the Nature of Explanation in 1943. He offered a pursuit of a Theory of Knowledge which would be contrastable like any other Natural Science. He was not completely successful in achieving this aim but he did, however, establish the rational bases upon which all the theories and models of systems of artificial behaviour have since been built. Craick offered a clear and powerful framework within which to express the acquisition, processing, storage, communication and use of knowledge (see figure 2). And third, the work of Warren McCulloch and Walter Pitts, A Logical Calculus of the Ideas Immanent in Nervous Activity, which was also published in 1943. They elaborated the concept of a “formal neuron”. Its response is, in fact, equivalent to a symbolic proposal with respect to the corresponding stimulus and which allows for a neural network to be considered as a logical system capable of handling symbols and elevating them to the level of the logic required for proposals. They came to the final conclusion that a network of formal neurons, with an effective infinite memory tape, can compute any number which is computable by a Turing Machine (see figure 3). Neurocybernetics and Artificial Intelligence 157 Figure 1. Fundamentals of homeostasis and reactive agents. Figure 2. Basis of Symbolic A.I. Agents and Robotics. 158 Roberto Moreno-Díaz In 1942 McCulloch met Norbert Wiener with their mutual friend Arturo Rosemblueth. The crucial paper for the emergence of Cybernetics was presented at the first Macy Foun- dation meeting in New York City: Behaviour, Purpose and Teleology, published the follow- ing year by Norbert Wiener, Arturo Rosemblueth and Julian Bigelow. Figure 3. Origins of Connectivism. Rosemblueth and McCulloch had reached with the Josiah Macy Foundation an important agreement to organize a yearly interdisciplinary meeting. Before they started, there was late in 1942 a meeting of engineers, physiologists and mathema- ticians at Princeton, referred by Wiener in the Introduction of his book of 1949 Cybernetics. There, McCulloch says, he met John von Neumann. The Macy Foundation Conferences started under the name Conferences“ on Circular, Causal and Feedback Mechanisms in Biological and Social Systems”, which was changed to “Conferences on Cybernetics” in 1949. The series of stimulating and constructive Conferences run until 1953. They established a new conception for treating living and non-living machines, which with more or less successes, fail- ures and redefinitions came to our days. It would be fruitful to dig into that remark- able source of ideas and inspiration, but only some of the Transactions are available. As Von Foerster said “The Conferences have became an oral tradition, a myth”. Today, the consequences of that myth can be found in McCulloch and other at- tendants’ essays. Many of the attendants to the Conferences may be considered the real foundations of Cybernetics. Some names: W. Ross Ashby, Y. Bar-Hillel, Julian Neurocybernetics and Artificial Intelligence 159 Bigelow, Jan Droogleever-Fortuyn, W. Grey Walter, Rafael Lorente de N´ o, Don- ald MacKay, Warren McCulloch (Chairman of the Conferences), J.M. Nielsen, F.S.C. Northrop, Linus Pauling, Antoine Remond, Arturo Rosemblueth, Claude Shannon, Heinz Von Foerster, John Von Neumann, Norbert Wiener. From that time, McCulloch was advisor and friend of the operational research pioneer Staf- ford Beer. From the Macy’s Conferences on, there were a number of crucial subjects and problems, raised and discussed in the many sessions. Among them, there were the concepts of regulation, homeostasis and goal directed activity, the transmission of signals and communication in machines and nervous systems, the raise of Neural Nets and Automata Theory. In what refers to nervous systems organization, the ideas of reverberating and closing loops to explain brain activity were established there. These ideas generated concepts and theories on circular causality in Econom- ics and in the polling of public opinion. Table 1. Neurocybernetics Subjects from the Macy’s Conferences (1950’s) * Regulation, homeostasis and goal directed activity * Transmission of signals and communication * Neural nets and automata theory * Closed loops in the central nervous system * “Circular causality” in economics and the polling of public opinion * Conflict between motives in psychiatry (heterarchy of values) * Reverberating and content addressable memories * Learning as changes in transition probabilities The analysis of conflict between motives in psychiatry led to the developing of con- cepts like heterarchyof values in mental processes. Also, the ideas of content ad- dressable memory, active or reverberating memories and the consideration of learn- ing as changes in transition probabilities among states, were inspired from Biology to became terms applicable to machines. In sum, a considerable and rich flow of new ideas and concepts to be applied both to machine and to living systems (see a summary in table 1). 160 Roberto Moreno-Díaz Neurocybernetics evolved with powerful input from the Theory of Communi- cation of Shannon and key figures in the field of Computer Science such as Von Neumann, in the case of the latter, with application to questions of computability, performability, capacity for reproduction and reliability of functioning. McCul- loch and Von Neumann were personal friends. McCulloch delighted a great deal in recounting the anecdote of how they began their work together on reliability of functioning, probabilistic logic and probabilistic computing. Table 2. Later Classical Neurocybernetics Problems Reliable computation in nets (McCulloch, von Neumann) Adaptive Systems and Learning (Ashby, von Foerster, Caianiello) McCulloch’s Programs I and II (Logical Synthesis: Neuronal Counterparts of Logical Machines) (“Automata” Synthesis of Nervous Structures) Connections Neurocybernetics-Artificial Intelligence McCulloch then held (in the Fifties) the chair of Psychiatry at the University of Chicago. One night, he, Von Neumann and some colleagues drunk too much whisky. McCulloch suddenly stopped the conversation dead and commented something on its effect: “The thresholds of neurones are now very, very low. Nev- ertheless, neurons are still computing reasonably reliably. What can there be in the brain, in its modular structure and links, which makes it such a reliable piece of machinery in spite of failure in threshold levels and components?” A magnificent piece of work called Agathe Tyche: The lucky reckoners offers a fair overview of muchof his philosophy with respect to ways of building reliable ma- chinery from unsafe components. The classic by Cowan called Reliable Computa- tion in the Presence of Noise and almost all of his later work on reliable computing was the result of McCulloch’s expansion of Von Neumann’s original concept. Master contributions in the 50’s and 60’s include among other Ross Ashby, Heinz von Foerster and Eduardo Caianiello. Ashby’s concept of homeostatic machines is fundamental for the development of mathematical Cybernetics, as well the ideas Neurocybernetics and Artificial Intelligence 161 developed in his classical book “Design for a Brain”. Von Foerster, was a physicist who became cybernetician after serving as secretary to the Macy’s Foundation Conferences and editor of the Transactions. His contributions on second order cybernetics or cybernetics of observing systems, are crucial to understand complex non trivial machines and systems. Third, there are Eduardo Caianiello’s neuronic and mnemonic equations for neural dynamics and for learning. Around 1965, some forty four years back, the office of McCulloch